GLM-4.5: Z.ai's Fully Open Release, Down to the FP8 Weights

What GLM-4.5 is

GLM-4.5 is Z.ai's open-weight model with a 128K token context window — recorded here as 131.1K, the same figure written out. What distinguishes the release is its completeness. Z.ai open-sourced the base models, the hybrid reasoning models, and FP8 versions of the hybrid reasoning models, for both GLM-4.5 and the smaller GLM-4.5-Air, all under the MIT licence with commercial use and secondary development explicitly permitted. Publishing base checkpoints alongside instruction-tuned ones is what makes a release useful to people doing their own post-training rather than only to people serving it, and shipping FP8 builds means the quantised version is the lab's own rather than a third party's approximation.

Benchmarks

Z.ai report an aggregate score of 63.2 across twelve industry-standard benchmarks, placing GLM-4.5 third among all proprietary and open-source models at the time of their evaluation, with GLM-4.5-Air at 59.8. The suite was selected by the vendor. On LMArena's text leaderboard GLM 4.5 has 1411 Elo (rank 153); by category, 1455 Elo in Coding (rank 153), 1411 Elo in Math (rank 145), 1445 Elo in Expert (rank 130) and 1372 Elo in Creative Writing (rank 156).

Fine-tuning hardware

Fine-tuning GLM-4.5 with LoRA requires 16 H100 GPUs at a batch size of 1 per GPU; GLM-4.5-Air needs 4.

Licence and price

MIT terms with explicit commercial and derivative rights, base checkpoints, hybrid reasoning variants and first-party FP8 builds. Around $0.60 per million input tokens through third-party providers, aggregated rather than first-party.

Architecture and modes

GLM-4.5 has 355 billion total parameters with 32 billion active per token. Z.ai describe the series as foundation models designed for intelligent agents, unifying reasoning, coding and agent capabilities. It is a hybrid reasoning model with a thinking mode for complex reasoning and tool use and a non-thinking mode for immediate responses; under vLLM and SGLang thinking is on by default and is turned off through the enable_thinking chat-template argument. Tool calls use the OpenAI-style tool description format. The weights are published on Hugging Face and ModelScope as GLM-4.5 (BF16), GLM-4.5-FP8 and GLM-4.5-Base.

Hardware for inference

Z.ai's minimum configuration for full-featured inference with SGLang is sixteen H100s or eight H200s in BF16, and eight H100s or four H200s in FP8. Running the full 128K context needs 32 H100s or 16 H200s in BF16, or 16 H100s or eight H200s in FP8. The configurations assume multi-token prediction with speculative decoding, no CPU offload, a batch size of at most eight, native FP8 hardware and more than 1TB of server memory. With the Swift framework on 96GB H20 GPUs, LoRA needs 16 GPUs and full supervised fine-tuning or reinforcement learning 128. On eight H100s under vLLM, Z.ai note that --cpu-offload-gb 16 may be needed.

On the Z.ai API

Z.ai's API documentation gives a 128K context length and a 96K maximum output, text in and text out, with thinking toggled by the thinking.type parameter and dynamic thinking on by default; function calling, streaming, context caching and JSON structured output are supported. Z.ai state pre-training on 15 trillion tokens, and list the twelve benchmarks behind its composite: MMLU Pro, AIME24, MATH 500, SciCode, GPQA, HLE, LiveCodeBench, SWE-Bench, Terminal-bench, TAU-Bench, BFCL v3 and BrowseComp. The documentation quotes API prices from $0.2 per million input tokens and $1.1 output for the series, and generation above 100 tokens per second in Z.ai's tests.

Official Resources